[deliverable/linux.git] / drivers / oprofile / buffer_sync.c

/**
 * @file buffer_sync.c
 *
 * @remark Copyright 2002 OProfile authors
 * @remark Read the file COPYING
 *
 * @author John Levon <levon@movementarian.org>
 *
 * This is the core of the buffer management. Each
 * CPU buffer is processed and entered into the
 * global event buffer. Such processing is necessary
 * in several circumstances, mentioned below.
 *
 * The processing does the job of converting the
 * transitory EIP value into a persistent dentry/offset
 * value that the profiler can record at its leisure.
 *
 * See fs/dcookies.c for a description of the dentry/offset
 * objects.
 */

#include <linux/mm.h>
#include <linux/workqueue.h>
#include <linux/notifier.h>
#include <linux/dcookies.h>
#include <linux/profile.h>
#include <linux/module.h>
#include <linux/fs.h>
#include <linux/oprofile.h>
#include <linux/sched.h>

#include "oprofile_stats.h"
#include "event_buffer.h"
#include "cpu_buffer.h"
#include "buffer_sync.h"

static LIST_HEAD(dying_tasks);
static LIST_HEAD(dead_tasks);
static cpumask_t marked_cpus = CPU_MASK_NONE;
static DEFINE_SPINLOCK(task_mortuary);
static void process_task_mortuary(void);


/* Take ownership of the task struct and place it on the
 * list for processing. Only after two full buffer syncs
 * does the task eventually get freed, because by then
 * we are sure we will not reference it again.
 * Can be invoked from softirq via RCU callback due to
 * call_rcu() of the task struct, hence the _irqsave.
 */
static int
task_free_notify(struct notifier_block *self, unsigned long val, void *data)
{
	unsigned long flags;
	struct task_struct *task = data;
	spin_lock_irqsave(&task_mortuary, flags);
	list_add(&task->tasks, &dying_tasks);
	spin_unlock_irqrestore(&task_mortuary, flags);
	return NOTIFY_OK;
}


/* The task is on its way out. A sync of the buffer means we can catch
 * any remaining samples for this task.
 */
static int
task_exit_notify(struct notifier_block *self, unsigned long val, void *data)
{
	/* To avoid latency problems, we only process the current CPU,
	 * hoping that most samples for the task are on this CPU
	 */
	sync_buffer(raw_smp_processor_id());
	return 0;
}


/* The task is about to try a do_munmap(). We peek at what it's going to
 * do, and if it's an executable region, process the samples first, so
 * we don't lose any. This does not have to be exact, it's a QoI issue
 * only.
 */
static int
munmap_notify(struct notifier_block *self, unsigned long val, void *data)
{
	unsigned long addr = (unsigned long)data;
	struct mm_struct *mm = current->mm;
	struct vm_area_struct *mpnt;

	down_read(&mm->mmap_sem);

	mpnt = find_vma(mm, addr);
	if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
		up_read(&mm->mmap_sem);
		/* To avoid latency problems, we only process the current CPU,
		 * hoping that most samples for the task are on this CPU
		 */
		sync_buffer(raw_smp_processor_id());
		return 0;
	}

	up_read(&mm->mmap_sem);
	return 0;
}


/* We need to be told about new modules so we don't attribute to a previously
 * loaded module, or drop the samples on the floor.
 */
static int
module_load_notify(struct notifier_block *self, unsigned long val, void *data)
{
#ifdef CONFIG_MODULES
	if (val != MODULE_STATE_COMING)
		return 0;

	/* FIXME: should we process all CPU buffers ? */
	mutex_lock(&buffer_mutex);
	add_event_entry(ESCAPE_CODE);
	add_event_entry(MODULE_LOADED_CODE);
	mutex_unlock(&buffer_mutex);
#endif
	return 0;
}


static struct notifier_block task_free_nb = {
	.notifier_call	= task_free_notify,
};

static struct notifier_block task_exit_nb = {
	.notifier_call	= task_exit_notify,
};

static struct notifier_block munmap_nb = {
	.notifier_call	= munmap_notify,
};

static struct notifier_block module_load_nb = {
	.notifier_call = module_load_notify,
};


static void end_sync(void)
{
	end_cpu_work();
	/* make sure we don't leak task structs */
	process_task_mortuary();
	process_task_mortuary();
}


int sync_start(void)
{
	int err;

	start_cpu_work();

	err = task_handoff_register(&task_free_nb);
	if (err)
		goto out1;
	err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
	if (err)
		goto out2;
	err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
	if (err)
		goto out3;
	err = register_module_notifier(&module_load_nb);
	if (err)
		goto out4;

out:
	return err;
out4:
	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
out3:
	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
out2:
	task_handoff_unregister(&task_free_nb);
out1:
	end_sync();
	goto out;
}


void sync_stop(void)
{
	unregister_module_notifier(&module_load_nb);
	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
	task_handoff_unregister(&task_free_nb);
	end_sync();
}


/* Optimisation. We can manage without taking the dcookie sem
 * because we cannot reach this code without at least one
 * dcookie user still being registered (namely, the reader
 * of the event buffer). */
static inline unsigned long fast_get_dcookie(struct path *path)
{
	unsigned long cookie;

	if (path->dentry->d_cookie)
		return (unsigned long)path->dentry;
	get_dcookie(path, &cookie);
	return cookie;
}


/* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
 * which corresponds loosely to "application name". This is
 * not strictly necessary but allows oprofile to associate
 * shared-library samples with particular applications
 */
static unsigned long get_exec_dcookie(struct mm_struct *mm)
{
	unsigned long cookie = NO_COOKIE;
	struct vm_area_struct *vma;

	if (!mm)
		goto out;

	for (vma = mm->mmap; vma; vma = vma->vm_next) {
		if (!vma->vm_file)
			continue;
		if (!(vma->vm_flags & VM_EXECUTABLE))
			continue;
		cookie = fast_get_dcookie(&vma->vm_file->f_path);
		break;
	}

out:
	return cookie;
}


/* Convert the EIP value of a sample into a persistent dentry/offset
 * pair that can then be added to the global event buffer. We make
 * sure to do this lookup before a mm->mmap modification happens so
 * we don't lose track.
 */
static unsigned long
lookup_dcookie(struct mm_struct *mm, unsigned long addr, off_t *offset)
{
	unsigned long cookie = NO_COOKIE;
	struct vm_area_struct *vma;

	for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {

		if (addr < vma->vm_start || addr >= vma->vm_end)
			continue;

		if (vma->vm_file) {
			cookie = fast_get_dcookie(&vma->vm_file->f_path);
			*offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
				vma->vm_start;
		} else {
			/* must be an anonymous map */
			*offset = addr;
		}

		break;
	}

	if (!vma)
		cookie = INVALID_COOKIE;

	return cookie;
}

static void increment_tail(struct oprofile_cpu_buffer *b)
{
	unsigned long new_tail = b->tail_pos + 1;

	rmb();

	if (new_tail < b->buffer_size)
		b->tail_pos = new_tail;
	else
		b->tail_pos = 0;
}

static unsigned long last_cookie = INVALID_COOKIE;

static void add_cpu_switch(int i)
{
	add_event_entry(ESCAPE_CODE);
	add_event_entry(CPU_SWITCH_CODE);
	add_event_entry(i);
	last_cookie = INVALID_COOKIE;
}

static void add_kernel_ctx_switch(unsigned int in_kernel)
{
	add_event_entry(ESCAPE_CODE);
	if (in_kernel)
		add_event_entry(KERNEL_ENTER_SWITCH_CODE);
	else
		add_event_entry(KERNEL_EXIT_SWITCH_CODE);
}

static void
add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
{
	add_event_entry(ESCAPE_CODE);
	add_event_entry(CTX_SWITCH_CODE);
	add_event_entry(task->pid);
	add_event_entry(cookie);
	/* Another code for daemon back-compat */
	add_event_entry(ESCAPE_CODE);
	add_event_entry(CTX_TGID_CODE);
	add_event_entry(task->tgid);
}


static void add_cookie_switch(unsigned long cookie)
{
	add_event_entry(ESCAPE_CODE);
	add_event_entry(COOKIE_SWITCH_CODE);
	add_event_entry(cookie);
}


static void add_trace_begin(void)
{
	add_event_entry(ESCAPE_CODE);
	add_event_entry(TRACE_BEGIN_CODE);
}


static void add_sample_entry(unsigned long offset, unsigned long event)
{
	add_event_entry(offset);
	add_event_entry(event);
}


static int add_us_sample(struct mm_struct *mm, struct op_sample *s)
{
	unsigned long cookie;
	off_t offset;

	cookie = lookup_dcookie(mm, s->eip, &offset);

	if (cookie == INVALID_COOKIE) {
		atomic_inc(&oprofile_stats.sample_lost_no_mapping);
		return 0;
	}

	if (cookie != last_cookie) {
		add_cookie_switch(cookie);
		last_cookie = cookie;
	}

	add_sample_entry(offset, s->event);

	return 1;
}


/* Add a sample to the global event buffer. If possible the
 * sample is converted into a persistent dentry/offset pair
 * for later lookup from userspace.
 */
static int
add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
{
	if (in_kernel) {
		add_sample_entry(s->eip, s->event);
		return 1;
	} else if (mm) {
		return add_us_sample(mm, s);
	} else {
		atomic_inc(&oprofile_stats.sample_lost_no_mm);
	}
	return 0;
}


static void release_mm(struct mm_struct *mm)
{
	if (!mm)
		return;
	up_read(&mm->mmap_sem);
	mmput(mm);
}


static struct mm_struct *take_tasks_mm(struct task_struct *task)
{
	struct mm_struct *mm = get_task_mm(task);
	if (mm)
		down_read(&mm->mmap_sem);
	return mm;
}


static inline int is_code(unsigned long val)
{
	return val == ESCAPE_CODE;
}


/* "acquire" as many cpu buffer slots as we can */
static unsigned long get_slots(struct oprofile_cpu_buffer *b)
{
	unsigned long head = b->head_pos;
	unsigned long tail = b->tail_pos;

	/*
	 * Subtle. This resets the persistent last_task
	 * and in_kernel values used for switching notes.
	 * BUT, there is a small window between reading
	 * head_pos, and this call, that means samples
	 * can appear at the new head position, but not
	 * be prefixed with the notes for switching
	 * kernel mode or a task switch. This small hole
	 * can lead to mis-attribution or samples where
	 * we don't know if it's in the kernel or not,
	 * at the start of an event buffer.
	 */
	cpu_buffer_reset(b);

	if (head >= tail)
		return head - tail;

	return head + (b->buffer_size - tail);
}


/* Move tasks along towards death. Any tasks on dead_tasks
 * will definitely have no remaining references in any
 * CPU buffers at this point, because we use two lists,
 * and to have reached the list, it must have gone through
 * one full sync already.
 */
static void process_task_mortuary(void)
{
	unsigned long flags;
	LIST_HEAD(local_dead_tasks);
	struct task_struct *task;
	struct task_struct *ttask;

	spin_lock_irqsave(&task_mortuary, flags);

	list_splice_init(&dead_tasks, &local_dead_tasks);
	list_splice_init(&dying_tasks, &dead_tasks);

	spin_unlock_irqrestore(&task_mortuary, flags);

	list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
		list_del(&task->tasks);
		free_task(task);
	}
}


static void mark_done(int cpu)
{
	int i;

	cpu_set(cpu, marked_cpus);

	for_each_online_cpu(i) {
		if (!cpu_isset(i, marked_cpus))
			return;
	}

	/* All CPUs have been processed at least once,
	 * we can process the mortuary once
	 */
	process_task_mortuary();

	cpus_clear(marked_cpus);
}


/* FIXME: this is not sufficient if we implement syscall barrier backtrace
 * traversal, the code switch to sb_sample_start at first kernel enter/exit
 * switch so we need a fifth state and some special handling in sync_buffer()
 */
typedef enum {
	sb_bt_ignore = -2,
	sb_buffer_start,
	sb_bt_start,
	sb_sample_start,
} sync_buffer_state;

/* Sync one of the CPU's buffers into the global event buffer.
 * Here we need to go through each batch of samples punctuated
 * by context switch notes, taking the task's mmap_sem and doing
 * lookup in task->mm->mmap to convert EIP into dcookie/offset
 * value.
 */
void sync_buffer(int cpu)
{
	struct oprofile_cpu_buffer *cpu_buf = &per_cpu(cpu_buffer, cpu);
	struct mm_struct *mm = NULL;
	struct task_struct *new;
	unsigned long cookie = 0;
	int in_kernel = 1;
	unsigned int i;
	sync_buffer_state state = sb_buffer_start;
	unsigned long available;

	mutex_lock(&buffer_mutex);

	add_cpu_switch(cpu);

	/* Remember, only we can modify tail_pos */

	available = get_slots(cpu_buf);

	for (i = 0; i < available; ++i) {
		struct op_sample *s = &cpu_buf->buffer[cpu_buf->tail_pos];

		if (is_code(s->eip)) {
			if (s->event <= CPU_IS_KERNEL) {
				/* kernel/userspace switch */
				in_kernel = s->event;
				if (state == sb_buffer_start)
					state = sb_sample_start;
				add_kernel_ctx_switch(s->event);
			} else if (s->event == CPU_TRACE_BEGIN) {
				state = sb_bt_start;
				add_trace_begin();
			} else {
				struct mm_struct *oldmm = mm;

				/* userspace context switch */
				new = (struct task_struct *)s->event;

				release_mm(oldmm);
				mm = take_tasks_mm(new);
				if (mm != oldmm)
					cookie = get_exec_dcookie(mm);
				add_user_ctx_switch(new, cookie);
			}
		} else if (state >= sb_bt_start &&
			   !add_sample(mm, s, in_kernel)) {
			if (state == sb_bt_start) {
				state = sb_bt_ignore;
				atomic_inc(&oprofile_stats.bt_lost_no_mapping);
			}
		}

		increment_tail(cpu_buf);
	}
	release_mm(mm);

	mark_done(cpu);

	mutex_unlock(&buffer_mutex);
}
Commit	Line	Data
1da177e4 LT	1	/**
	2	* @file buffer_sync.c
	3	*
	4	* @remark Copyright 2002 OProfile authors
	5	* @remark Read the file COPYING
	6	*
	7	* @author John Levon <levon@movementarian.org>
	8	*
	9	* This is the core of the buffer management. Each
	10	* CPU buffer is processed and entered into the
	11	* global event buffer. Such processing is necessary
	12	* in several circumstances, mentioned below.
	13	*
	14	* The processing does the job of converting the
	15	* transitory EIP value into a persistent dentry/offset
	16	* value that the profiler can record at its leisure.
	17	*
	18	* See fs/dcookies.c for a description of the dentry/offset
	19	* objects.
	20	*/
	21
	22	#include <linux/mm.h>
	23	#include <linux/workqueue.h>
	24	#include <linux/notifier.h>
	25	#include <linux/dcookies.h>
	26	#include <linux/profile.h>
	27	#include <linux/module.h>
	28	#include <linux/fs.h>
1474855d	29	#include <linux/oprofile.h>
e8edc6e0	30	#include <linux/sched.h>
1474855d	31
1da177e4 LT	32	#include "oprofile_stats.h"
	33	#include "event_buffer.h"
	34	#include "cpu_buffer.h"
	35	#include "buffer_sync.h"
73185e0a	36
1da177e4 LT	37	static LIST_HEAD(dying_tasks);
	38	static LIST_HEAD(dead_tasks);
	39	static cpumask_t marked_cpus = CPU_MASK_NONE;
	40	static DEFINE_SPINLOCK(task_mortuary);
	41	static void process_task_mortuary(void);
	42
	43
	44	/* Take ownership of the task struct and place it on the
	45	* list for processing. Only after two full buffer syncs
	46	* does the task eventually get freed, because by then
	47	* we are sure we will not reference it again.
4369ef3c PM	48	* Can be invoked from softirq via RCU callback due to
4369ef3c PM	49	* call_rcu() of the task struct, hence the _irqsave.
1da177e4	50	*/
73185e0a RR	51	static int
73185e0a RR	52	task_free_notify(struct notifier_block self, unsigned long val, void data)
1da177e4	53	{
4369ef3c	54	unsigned long flags;
73185e0a	55	struct task_struct *task = data;
4369ef3c	56	spin_lock_irqsave(&task_mortuary, flags);
1da177e4	57	list_add(&task->tasks, &dying_tasks);
4369ef3c	58	spin_unlock_irqrestore(&task_mortuary, flags);
1da177e4 LT	59	return NOTIFY_OK;
	60	}
	61
	62
	63	/* The task is on its way out. A sync of the buffer means we can catch
	64	* any remaining samples for this task.
	65	*/
73185e0a RR	66	static int
73185e0a RR	67	task_exit_notify(struct notifier_block self, unsigned long val, void data)
1da177e4 LT	68	{
	69	/* To avoid latency problems, we only process the current CPU,
	70	* hoping that most samples for the task are on this CPU
	71	*/
39c715b7	72	sync_buffer(raw_smp_processor_id());
73185e0a	73	return 0;
1da177e4 LT	74	}
	75
	76
	77	/* The task is about to try a do_munmap(). We peek at what it's going to
	78	* do, and if it's an executable region, process the samples first, so
	79	* we don't lose any. This does not have to be exact, it's a QoI issue
	80	* only.
	81	*/
73185e0a RR	82	static int
73185e0a RR	83	munmap_notify(struct notifier_block self, unsigned long val, void data)
1da177e4 LT	84	{
1da177e4 LT	85	unsigned long addr = (unsigned long)data;
73185e0a RR	86	struct mm_struct *mm = current->mm;
73185e0a RR	87	struct vm_area_struct *mpnt;
1da177e4 LT	88
	89	down_read(&mm->mmap_sem);
	90
	91	mpnt = find_vma(mm, addr);
	92	if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
	93	up_read(&mm->mmap_sem);
	94	/* To avoid latency problems, we only process the current CPU,
	95	* hoping that most samples for the task are on this CPU
	96	*/
39c715b7	97	sync_buffer(raw_smp_processor_id());
1da177e4 LT	98	return 0;
	99	}
	100
	101	up_read(&mm->mmap_sem);
	102	return 0;
	103	}
	104
73185e0a	105
1da177e4 LT	106	/* We need to be told about new modules so we don't attribute to a previously
	107	* loaded module, or drop the samples on the floor.
	108	*/
73185e0a RR	109	static int
73185e0a RR	110	module_load_notify(struct notifier_block self, unsigned long val, void data)
1da177e4 LT	111	{
	112	#ifdef CONFIG_MODULES
	113	if (val != MODULE_STATE_COMING)
	114	return 0;
	115
	116	/* FIXME: should we process all CPU buffers ? */
59cc185a	117	mutex_lock(&buffer_mutex);
1da177e4 LT	118	add_event_entry(ESCAPE_CODE);
1da177e4 LT	119	add_event_entry(MODULE_LOADED_CODE);
59cc185a	120	mutex_unlock(&buffer_mutex);
1da177e4 LT	121	#endif
	122	return 0;
	123	}
	124
73185e0a	125
1da177e4 LT	126	static struct notifier_block task_free_nb = {
	127	.notifier_call = task_free_notify,
	128	};
	129
	130	static struct notifier_block task_exit_nb = {
	131	.notifier_call = task_exit_notify,
	132	};
	133
	134	static struct notifier_block munmap_nb = {
	135	.notifier_call = munmap_notify,
	136	};
	137
	138	static struct notifier_block module_load_nb = {
	139	.notifier_call = module_load_notify,
	140	};
	141
73185e0a	142
1da177e4 LT	143	static void end_sync(void)
	144	{
	145	end_cpu_work();
	146	/* make sure we don't leak task structs */
	147	process_task_mortuary();
	148	process_task_mortuary();
	149	}
	150
	151
	152	int sync_start(void)
	153	{
	154	int err;
	155
	156	start_cpu_work();
	157
	158	err = task_handoff_register(&task_free_nb);
	159	if (err)
	160	goto out1;
	161	err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
	162	if (err)
	163	goto out2;
	164	err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
	165	if (err)
	166	goto out3;
	167	err = register_module_notifier(&module_load_nb);
	168	if (err)
	169	goto out4;
	170
	171	out:
	172	return err;
	173	out4:
	174	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
	175	out3:
	176	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
	177	out2:
	178	task_handoff_unregister(&task_free_nb);
	179	out1:
	180	end_sync();
	181	goto out;
	182	}
	183
	184
	185	void sync_stop(void)
	186	{
	187	unregister_module_notifier(&module_load_nb);
	188	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
	189	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
	190	task_handoff_unregister(&task_free_nb);
	191	end_sync();
	192	}
	193
448678a0	194
1da177e4 LT	195	/* Optimisation. We can manage without taking the dcookie sem
	196	* because we cannot reach this code without at least one
	197	* dcookie user still being registered (namely, the reader
	198	* of the event buffer). */
448678a0	199	static inline unsigned long fast_get_dcookie(struct path *path)
1da177e4 LT	200	{
1da177e4 LT	201	unsigned long cookie;
448678a0 JB	202
	203	if (path->dentry->d_cookie)
	204	return (unsigned long)path->dentry;
	205	get_dcookie(path, &cookie);
1da177e4 LT	206	return cookie;
	207	}
	208
448678a0	209
1da177e4 LT	210	/* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
	211	* which corresponds loosely to "application name". This is
	212	* not strictly necessary but allows oprofile to associate
	213	* shared-library samples with particular applications
	214	*/
73185e0a	215	static unsigned long get_exec_dcookie(struct mm_struct *mm)
1da177e4	216	{
0c0a400d	217	unsigned long cookie = NO_COOKIE;
73185e0a RR	218	struct vm_area_struct *vma;
73185e0a RR	219
1da177e4 LT	220	if (!mm)
1da177e4 LT	221	goto out;
73185e0a	222
1da177e4 LT	223	for (vma = mm->mmap; vma; vma = vma->vm_next) {
	224	if (!vma->vm_file)
	225	continue;
	226	if (!(vma->vm_flags & VM_EXECUTABLE))
	227	continue;
448678a0	228	cookie = fast_get_dcookie(&vma->vm_file->f_path);
1da177e4 LT	229	break;
	230	}
	231
	232	out:
	233	return cookie;
	234	}
	235
	236
	237	/* Convert the EIP value of a sample into a persistent dentry/offset
	238	* pair that can then be added to the global event buffer. We make
	239	* sure to do this lookup before a mm->mmap modification happens so
	240	* we don't lose track.
	241	*/
73185e0a RR	242	static unsigned long
73185e0a RR	243	lookup_dcookie(struct mm_struct mm, unsigned long addr, off_t offset)
1da177e4	244	{
0c0a400d	245	unsigned long cookie = NO_COOKIE;
73185e0a	246	struct vm_area_struct *vma;
1da177e4 LT	247
1da177e4 LT	248	for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
73185e0a	249
1da177e4 LT	250	if (addr < vma->vm_start \|\| addr >= vma->vm_end)
	251	continue;
	252
0c0a400d	253	if (vma->vm_file) {
448678a0	254	cookie = fast_get_dcookie(&vma->vm_file->f_path);
0c0a400d JL	255	*offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
	256	vma->vm_start;
	257	} else {
	258	/* must be an anonymous map */
	259	*offset = addr;
	260	}
	261
1da177e4 LT	262	break;
	263	}
	264
0c0a400d JL	265	if (!vma)
	266	cookie = INVALID_COOKIE;
	267
1da177e4 LT	268	return cookie;
	269	}
	270
5e11f98d RR	271	static void increment_tail(struct oprofile_cpu_buffer *b)
	272	{
	273	unsigned long new_tail = b->tail_pos + 1;
	274
	275	rmb();
	276
	277	if (new_tail < b->buffer_size)
	278	b->tail_pos = new_tail;
	279	else
	280	b->tail_pos = 0;
	281	}
1da177e4	282
0c0a400d	283	static unsigned long last_cookie = INVALID_COOKIE;
73185e0a	284
1da177e4 LT	285	static void add_cpu_switch(int i)
	286	{
	287	add_event_entry(ESCAPE_CODE);
	288	add_event_entry(CPU_SWITCH_CODE);
	289	add_event_entry(i);
0c0a400d	290	last_cookie = INVALID_COOKIE;
1da177e4 LT	291	}
	292
	293	static void add_kernel_ctx_switch(unsigned int in_kernel)
	294	{
	295	add_event_entry(ESCAPE_CODE);
	296	if (in_kernel)
73185e0a	297	add_event_entry(KERNEL_ENTER_SWITCH_CODE);
1da177e4	298	else
73185e0a	299	add_event_entry(KERNEL_EXIT_SWITCH_CODE);
1da177e4	300	}
73185e0a	301
1da177e4	302	static void
73185e0a	303	add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
1da177e4 LT	304	{
1da177e4 LT	305	add_event_entry(ESCAPE_CODE);
73185e0a	306	add_event_entry(CTX_SWITCH_CODE);
1da177e4 LT	307	add_event_entry(task->pid);
	308	add_event_entry(cookie);
	309	/* Another code for daemon back-compat */
	310	add_event_entry(ESCAPE_CODE);
	311	add_event_entry(CTX_TGID_CODE);
	312	add_event_entry(task->tgid);
	313	}
	314
73185e0a	315
1da177e4 LT	316	static void add_cookie_switch(unsigned long cookie)
	317	{
	318	add_event_entry(ESCAPE_CODE);
	319	add_event_entry(COOKIE_SWITCH_CODE);
	320	add_event_entry(cookie);
	321	}
	322
73185e0a	323
1da177e4 LT	324	static void add_trace_begin(void)
	325	{
	326	add_event_entry(ESCAPE_CODE);
	327	add_event_entry(TRACE_BEGIN_CODE);
	328	}
	329
	330
	331	static void add_sample_entry(unsigned long offset, unsigned long event)
	332	{
	333	add_event_entry(offset);
	334	add_event_entry(event);
	335	}
	336
	337
73185e0a	338	static int add_us_sample(struct mm_struct mm, struct op_sample s)
1da177e4 LT	339	{
	340	unsigned long cookie;
	341	off_t offset;
73185e0a RR	342
	343	cookie = lookup_dcookie(mm, s->eip, &offset);
	344
0c0a400d	345	if (cookie == INVALID_COOKIE) {
1da177e4 LT	346	atomic_inc(&oprofile_stats.sample_lost_no_mapping);
	347	return 0;
	348	}
	349
	350	if (cookie != last_cookie) {
	351	add_cookie_switch(cookie);
	352	last_cookie = cookie;
	353	}
	354
	355	add_sample_entry(offset, s->event);
	356
	357	return 1;
	358	}
	359
73185e0a	360
1da177e4 LT	361	/* Add a sample to the global event buffer. If possible the
	362	* sample is converted into a persistent dentry/offset pair
	363	* for later lookup from userspace.
	364	*/
	365	static int
73185e0a	366	add_sample(struct mm_struct mm, struct op_sample s, int in_kernel)
1da177e4 LT	367	{
	368	if (in_kernel) {
	369	add_sample_entry(s->eip, s->event);
	370	return 1;
	371	} else if (mm) {
	372	return add_us_sample(mm, s);
	373	} else {
	374	atomic_inc(&oprofile_stats.sample_lost_no_mm);
	375	}
	376	return 0;
	377	}
1da177e4	378
73185e0a RR	379
73185e0a RR	380	static void release_mm(struct mm_struct *mm)
1da177e4 LT	381	{
	382	if (!mm)
	383	return;
	384	up_read(&mm->mmap_sem);
	385	mmput(mm);
	386	}
	387
	388
73185e0a	389	static struct mm_struct take_tasks_mm(struct task_struct task)
1da177e4	390	{
73185e0a	391	struct mm_struct *mm = get_task_mm(task);
1da177e4 LT	392	if (mm)
	393	down_read(&mm->mmap_sem);
	394	return mm;
	395	}
	396
	397
	398	static inline int is_code(unsigned long val)
	399	{
	400	return val == ESCAPE_CODE;
	401	}
73185e0a	402
1da177e4 LT	403
1da177e4 LT	404	/* "acquire" as many cpu buffer slots as we can */
73185e0a	405	static unsigned long get_slots(struct oprofile_cpu_buffer *b)
1da177e4 LT	406	{
	407	unsigned long head = b->head_pos;
	408	unsigned long tail = b->tail_pos;
	409
	410	/*
	411	* Subtle. This resets the persistent last_task
	412	* and in_kernel values used for switching notes.
	413	* BUT, there is a small window between reading
	414	* head_pos, and this call, that means samples
	415	* can appear at the new head position, but not
	416	* be prefixed with the notes for switching
	417	* kernel mode or a task switch. This small hole
	418	* can lead to mis-attribution or samples where
	419	* we don't know if it's in the kernel or not,
	420	* at the start of an event buffer.
	421	*/
	422	cpu_buffer_reset(b);
	423
	424	if (head >= tail)
	425	return head - tail;
	426
	427	return head + (b->buffer_size - tail);
	428	}
	429
	430
1da177e4 LT	431	/* Move tasks along towards death. Any tasks on dead_tasks
	432	* will definitely have no remaining references in any
	433	* CPU buffers at this point, because we use two lists,
	434	* and to have reached the list, it must have gone through
	435	* one full sync already.
	436	*/
	437	static void process_task_mortuary(void)
	438	{
4369ef3c PM	439	unsigned long flags;
4369ef3c PM	440	LIST_HEAD(local_dead_tasks);
73185e0a RR	441	struct task_struct *task;
73185e0a RR	442	struct task_struct *ttask;
1da177e4	443
4369ef3c	444	spin_lock_irqsave(&task_mortuary, flags);
1da177e4	445
4369ef3c PM	446	list_splice_init(&dead_tasks, &local_dead_tasks);
4369ef3c PM	447	list_splice_init(&dying_tasks, &dead_tasks);
1da177e4	448
4369ef3c PM	449	spin_unlock_irqrestore(&task_mortuary, flags);
	450
	451	list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
1da177e4	452	list_del(&task->tasks);
4369ef3c	453	free_task(task);
1da177e4	454	}
1da177e4 LT	455	}
	456
	457
	458	static void mark_done(int cpu)
	459	{
	460	int i;
	461
	462	cpu_set(cpu, marked_cpus);
	463
	464	for_each_online_cpu(i) {
	465	if (!cpu_isset(i, marked_cpus))
	466	return;
	467	}
	468
	469	/* All CPUs have been processed at least once,
	470	* we can process the mortuary once
	471	*/
	472	process_task_mortuary();
	473
	474	cpus_clear(marked_cpus);
	475	}
	476
	477
	478	/* FIXME: this is not sufficient if we implement syscall barrier backtrace
	479	* traversal, the code switch to sb_sample_start at first kernel enter/exit
	480	* switch so we need a fifth state and some special handling in sync_buffer()
	481	*/
	482	typedef enum {
	483	sb_bt_ignore = -2,
	484	sb_buffer_start,
	485	sb_bt_start,
	486	sb_sample_start,
	487	} sync_buffer_state;
	488
	489	/* Sync one of the CPU's buffers into the global event buffer.
	490	* Here we need to go through each batch of samples punctuated
	491	* by context switch notes, taking the task's mmap_sem and doing
	492	* lookup in task->mm->mmap to convert EIP into dcookie/offset
	493	* value.
	494	*/
	495	void sync_buffer(int cpu)
	496	{
608dfddd	497	struct oprofile_cpu_buffer *cpu_buf = &per_cpu(cpu_buffer, cpu);
1da177e4	498	struct mm_struct *mm = NULL;
73185e0a	499	struct task_struct *new;
1da177e4 LT	500	unsigned long cookie = 0;
	501	int in_kernel = 1;
	502	unsigned int i;
	503	sync_buffer_state state = sb_buffer_start;
	504	unsigned long available;
	505
59cc185a	506	mutex_lock(&buffer_mutex);
73185e0a	507
1da177e4 LT	508	add_cpu_switch(cpu);
	509
	510	/* Remember, only we can modify tail_pos */
	511
	512	available = get_slots(cpu_buf);
	513
	514	for (i = 0; i < available; ++i) {
73185e0a RR	515	struct op_sample *s = &cpu_buf->buffer[cpu_buf->tail_pos];
73185e0a RR	516
1da177e4 LT	517	if (is_code(s->eip)) {
	518	if (s->event <= CPU_IS_KERNEL) {
	519	/* kernel/userspace switch */
	520	in_kernel = s->event;
	521	if (state == sb_buffer_start)
	522	state = sb_sample_start;
	523	add_kernel_ctx_switch(s->event);
	524	} else if (s->event == CPU_TRACE_BEGIN) {
	525	state = sb_bt_start;
	526	add_trace_begin();
	527	} else {
73185e0a	528	struct mm_struct *oldmm = mm;
1da177e4 LT	529
	530	/* userspace context switch */
	531	new = (struct task_struct *)s->event;
	532
	533	release_mm(oldmm);
	534	mm = take_tasks_mm(new);
	535	if (mm != oldmm)
	536	cookie = get_exec_dcookie(mm);
	537	add_user_ctx_switch(new, cookie);
	538	}
73185e0a RR	539	} else if (state >= sb_bt_start &&
	540	!add_sample(mm, s, in_kernel)) {
	541	if (state == sb_bt_start) {
	542	state = sb_bt_ignore;
	543	atomic_inc(&oprofile_stats.bt_lost_no_mapping);
1da177e4 LT	544	}
	545	}
	546
	547	increment_tail(cpu_buf);
	548	}
	549	release_mm(mm);
	550
	551	mark_done(cpu);
	552
59cc185a	553	mutex_unlock(&buffer_mutex);
1da177e4	554	}